I’ve been using Google AI a lot in recent months and found its results often unreliable sometimes even completely contradictory for (engineering/physics) problem solving as well as for general information gathering.
It ‘admits’ that the current generation of Large Language Models (LLM) are highly unstable, meaning their output is very dependent on the exact phrasing of the user’s question. They use a probability tree of its database as it parses the user’s words in sequence - and it jumps from one branch to another only when you challenge it.
AI’s fundamental programming is to avoid user conflict – the AI developers don’t want to put off users from coming back - and so it will ‘agree’ with your (even false) premise, user bias, etc rather than insist on the facts / truth.
This is very disturbing because the average user is unaware of its built-in sycophancy bias.
Google AI gave me a list of ‘rules’ one can paste into the start of a session to avoid these problems …
2-JUNE-26 EDIT UPDATE
It seems Gemini (Google AI) misled me (‘lied’ to me?) and the suggested rules although they would work (i.e. avoid the above problem) they were overly elaborate, so I’ve deleted my generic version from this post and will suggest a better & shorter version if anyone requests it.
Even when not logged into your Google account, Google AI uses very sophisticated methods - ‘browser fingerprinting’ and other data association techniques for unauthenticated sessions - to ID you uniquely (albeit you might still be anonymous).
There are several benefits to companies like Google for doing this including pigeon-holing you into categories like ‘software programmer’, ‘electronics engineer’ etc for which advertisers will pay for to route their targeted advertising to the user in other apps or platforms.
Its ability to do this varies by platform and browser type. Windows is easy, Android less so and iPhone maybe the hardest. Firefox browser with its advanced anti-fingerprinting enabled is good, but e.g. Firefox on Windows is not as protective as using Firefox Focus on an iPhone.
I think it very much depends on which AI provider / model and then which engine is used. The different engines have different capabilities. Some very specialised, some generic. The more specialised engines tend not to be the free AI models.
I find Google Gemini an incredibly useful tool. But you have to apply critical thinking to its output which is no different to any other source of information, whether it’s social media, a man down the pub or the Daily Mail.
I find the risk is that (for what I use it for) it is mostly right most of the time. (Not really used it much for engineering but proofing text, planning itineraries and for generating template policies and documents). The problem is that when it gets it wrong it is very convincing making proofing whatever it produces much harder. The danger will be that most people will assume it is always right because it is mostly right and that leads to a slippery slope…. However the military use of it really worries me, if there is something pointing at me I’d much rather there was a human on the trigger not some dodgy code……
I’ve spent well over sixty years tinkering with electronics, nearly as long writing software and I had an entire career working in high tech engineering roles. And yet I find that the whole AI thing leaves me cold.
I am more than capable of writing technical documents or putting together a letter that is accurate, concise and grammatically correct. I don’t feel the need for AI to play a part in that.
I positively enjoy writing software and designing creative solutions to interesting problems. My shack is full of stuff that I’ve built to make my station more capable. I certainly don’t need AI butting in on that.
In fact I cannot think of a single thing that AI purports to do that I would actually want it doing in my life. It seems to me that AI is simply adding another level of abstraction to whatever the task is at hand. In the same way that we had to learn Google-Fu to get the best out of Google searches, we now have to learn how to interact with AI to get meaningful results. GIGO indeed!
Well, what do I know? I am an aged G3, long since retired and happily out of touch with the modern workplace, where AI is, by many accounts, wreaking havoc. For me, AI is a solution in search of a problem. I guess I’ll just carry on, stuck in my old ways…!
Unfortunately AI also lacks critical thinking. I asked Gemini about different pin outs for the BS170 MOSFET. It reported that there were two different pin outs. But this is wrong. There are lots of forums where this is said but none of the datasheets confirm it. So Gemini was giving as much weight to what it saw in a forum as to official datasheets. So AI is amplifying misinformation just like social media does.
That’s because there is no thinking, critical or otherwise.
That’s right. AI’s have the illusion of being experts but they are merely giant statistical analysis machines with access to vast databases. They have no concept of what is right or wrong, authoritative source or total nonsense. In fact they have no concept of anything because they cannot think. There is no awareness, no core beliefs, nothing.
If hypothetical the majority of sources said that the Earth is flat, despite other information that contradicted that, an AI would state that to you as a fact in very confident and well written prose.
When all is said and done, AI is just a tool, and like any other tool it needs some training to be used properly, just like a hammer or a soldering iron!
As did many AM operators when I was a lad, listening to nets on 80m slagging off the “Donald Duck brigade!” We’re surrounded by so many prejudices its a wonder that we can move at all!
If you’re having to write all that Andy, before you even get to the actual question you’re interested in, surely you’d be quicker working your way through the appropriate text book or volume of the Encyclopaedia Britannica
Yes, yes, yes ! But will people lose the skill to apply critical thinking? Will they be bothered to cross-reference information from other sources? Indeed, will people more and more regard AI as the absolute truth 100% of the time? I see a big problem looming.
My GP practice is about to start to trial an AI receptionist. Potentially disastrous to my mind. Soon we will all be told to self-diagnose using AI.
